An efficient intelligent welding system

Abstract

The application provides a kind of efficient intelligent welding system, including multi-source perception module, intelligent prediction module and dynamic execution module, the multi-source perception module is with high-precision initialization by data fusion with intelligent prediction module, intelligent prediction module and dynamic execution module realize high-frequency regulation by closed-loop control, dynamic execution module and intelligent prediction module realize model self-evolution by real-time feedback.The application improves the overall performance of the welding process significantly through the collaborative design of multi-source perception, intelligent prediction and dynamic execution module, and the optimized preparation method.The system has self-evolution ability and can continuously optimize the model and strategy according to real-time feedback to maintain long-term operation stability.It improves welding efficiency and product quality, enhances the adaptability to complex workpieces and diversified production scenarios, reduces manual intervention and resource waste, provides efficient, reliable and green welding solutions for modern manufacturing, and shows significant technological progress and practical value.

Description

Technical Field

[0001] This invention relates to the field of welding, specifically to a highly efficient intelligent welding system. Background Technology

[0002] Welding technology, as a core process in modern manufacturing, is widely used in automotive, aerospace, and shipbuilding industries. With the increasing level of industrial automation, intelligent welding systems are gradually replacing traditional manual welding to improve production efficiency and product quality. However, existing welding technologies still have several technical limitations that restrict their application in complex working conditions and high-precision scenarios.

[0003] First, existing welding systems have shortcomings in terms of environmental adaptability and initialization accuracy. Traditional welding equipment often relies on a single optical sensor or manual parameter settings to determine the weld position and initial state, making it difficult to cope with interference from highly reflective materials or extreme environments. This results in large initial errors, typically exceeding ±0.1mm, affecting the stability of subsequent welding quality. Furthermore, existing technologies lack the ability to fuse multi-source data, failing to comprehensively capture the dynamic characteristics of the welding process.

[0004] Secondly, existing intelligent welding systems have significant limitations in predictive capabilities and parameter optimization. Traditional methods often employ fixed parameters or simple PID control, lacking long-term predictive capabilities for weld quality, energy consumption, and defect probability. The prediction window is typically no more than 5 minutes, and the accuracy is low. This makes it difficult for the system to plan and optimize parameters in advance, especially in scenarios with complex workpieces or multiple objectives, where welding efficiency and consistency are limited.

[0005] Secondly, existing welding systems suffer from insufficient adjustment response speed and closed-loop control capabilities. Traditional equipment typically employs open-loop control or low-frequency feedback regulation with an adjustment cycle of 10-30 seconds, failing to respond in real time to instantaneous changes during the welding process, such as molten pool fluctuations and spatter, resulting in low weld consistency (approximately 95%-97%) and a high defect rate. Furthermore, existing technologies lack precise control over the heat-affected zone, and overheating issues frequently impact workpiece performance.

[0006] Finally, existing welding systems generally lack adaptive learning capabilities. Traditional methods rely on preset parameters or offline adjustments, failing to dynamically optimize models and strategies based on real-time feedback. This results in significant quality fluctuations during long-term operation, making it difficult to meet the industrial demands for high-precision, continuous production.

[0007] In summary, existing technologies have significant shortcomings in terms of environmental adaptability, prediction accuracy, adjustment speed, and long-term stability, making it difficult to meet the requirements of modern manufacturing for efficient, high-quality, and intelligent welding. Summary of the Invention

[0008] The purpose of this invention is to provide an intelligent welding system and method that can achieve high-precision initialization, ultra-long prediction window, high-frequency closed-loop adjustment and self-evolutionary optimization, so as to improve welding efficiency, consistency and environmental adaptability, while reducing defect rate and energy consumption.

[0009] To achieve the above objectives, the present invention proposes the following technical solution: a high-efficiency intelligent welding system, characterized in that it comprises the following modules:

[0010] The multi-source sensing module includes a dual-wavelength laser vision sensor, a multi-band acoustic sensor, an infrared thermal imaging array, and an environmental adaptation unit, which are used to acquire weld position, arc sound signal, temperature distribution, and environmental parameters in real time.

[0011] The intelligent prediction module, including an edge computing accelerator and an embedded industrial computer, incorporates a prediction model that fuses Transformer and Long Short-Term Memory networks with a reinforcement learning optimization algorithm to predict weld quality, energy consumption, and defect probability, and generate optimized parameters.

[0012] The dynamic execution module includes a six-axis welding robot, a two-degree-of-freedom welding torch adjustment mechanism, an adaptive power control unit, and a microchannel cooling system, which are used to dynamically adjust the welding path, power parameters, and cooling intensity based on prediction results.

[0013] Among them, the multi-source sensing module and the intelligent prediction module achieve high-precision initialization through data fusion, the intelligent prediction module and the dynamic execution module achieve high-frequency adjustment through closed-loop control, and the dynamic execution module and the intelligent prediction module achieve model self-evolution through real-time feedback.

[0014] Furthermore, in this invention, the dynamic execution module includes:

[0015] The six-axis welding robot includes a six-axis robotic arm driven by six servo motors, a force feedback sensor mounted at the end effector, an embedded motion controller, and a high-precision optical encoder. The force feedback sensor has an accuracy of ±0.1N, and the encoder has a resolution of 0.01°, which are used to dynamically adjust the welding path based on predicted parameters.

[0016] The dual-degree-of-freedom welding torch adjustment mechanism includes a dual-axis servo motor, a linear guide rail, an angle sensor, and a servo driver. The dual-axis servo motor adjusts the height and tilt angle with an adjustment cycle of 0.5 seconds to ensure weld consistency.

[0017] The adaptive power control unit includes an IGBT inverter power supply, a pulse generator, a current sensor, and a control circuit. The inverter power supply has an output range of 50-300A, the pulse generator supports 10-500Hz modulation, and the current sensor has an accuracy of ±0.5A. It is used to dynamically adjust the current and pulse frequency to optimize the stability of the molten pool.

[0018] The microchannel cooling system includes a microchannel nozzle, a solenoid valve, a gas flow meter, and a pressure regulator. The microchannel nozzle has an orifice diameter of 0.1 mm, the solenoid valve adjusts the flow rate range to 0-10 L / min, the flow meter has an accuracy of ±0.1 L / min, and the pressure regulator maintains a pressure of 0.2-0.5 MPa to control the heat-affected zone.

[0019] Furthermore, in this invention, the force feedback sensor and motion controller of the six-axis welding robot work together to optimize the path by sensing the workpiece contact force in real time and running a reinforcement learning algorithm; the dual-axis servo motor and angle sensor of the dual-degree-of-freedom welding torch adjustment mechanism achieve closed-loop adjustment through a servo driver; the pulse generator and current sensor of the adaptive power control unit dynamically adjust the pulse waveform through a control circuit; and the solenoid valve and gas flow meter of the microchannel cooling system precisely adjust the cooling flow rate through feedback control.

[0020] Furthermore, in this invention, the six-axis welding robot of the dynamic execution module adjusts its path according to the optimized parameters output by the intelligent prediction module, and the dual-degree-of-freedom welding torch adjustment mechanism, the adaptive power control unit, and the microchannel cooling system adjust the angle, pulse frequency, and cooling flow rate according to the fuzzy control output.

[0021] According to the above-mentioned efficient intelligent welding system, the execution steps are as follows:

[0022] Step 1: Multidimensional initialization and environmental adaptive calibration. Initial data is collected using the multi-source sensing module, and the Bayesian probability model is run through the industrial computer in the intelligent prediction module to calibrate environmental disturbances and generate a high-precision initial state vector.

[0023] Step 2: AI-driven multi-objective prediction and optimization planning. The edge computing accelerator in the intelligent prediction module is used to run the Transformer model to predict welding trends. The industrial computer performs reinforcement learning to optimize parameters and generate a multi-objective planning scheme.

[0024] Step 3: Adaptive closed-loop welding execution. The dynamic execution module performs welding according to the predicted parameters, the multi-source sensing module monitors in real time, and the intelligent prediction module runs Kalman filtering and fuzzy control to dynamically adjust the execution parameters.

[0025] Step 4: Real-time feedback and AI model self-evolution. The multi-source perception module provides actual data, and the intelligent prediction module updates the Transformer model and RL strategy through edge computing accelerators and industrial computers to improve long-term performance.

[0026] Furthermore, in this invention, the specific process of step 1 is as follows: the dual-wavelength laser vision sensor of the multi-source sensing module scans the workpiece surface to generate a three-dimensional point cloud of the weld position x; the multi-band acoustic sensor records the initial arc sound frequency f through a microphone array; the infrared thermal imaging array measures the workpiece temperature distribution T to generate a two-dimensional thermal map; the environmental adaptation unit includes a humidity sensor and a pressure sensor to collect environmental parameters E, which include humidity h and pressure v; the prediction module includes an embedded industrial computer that runs a Bayesian algorithm, stores historical data, and calculates the probability. The calculation formula is:

[0027] (P(S|E)): Posterior probability, representing the welding state (S) given environmental parameters (E); (P(E|S)): Likelihood probability, generated from sensor data; (P(S)): Prior probability, based on historical welding data; (P(E)): Marginal probability of environmental parameters.

[0028] A dual-wavelength laser vision sensor scans the workpiece and outputs point cloud data; an acoustic sensor records the arc sound spectrum; an infrared thermal imaging array generates a thermal map; an environmental adaptation unit measures humidity and air pressure; an industrial computer receives this data, calculates a likelihood model based on sensor accuracy and historical database, outputs the posterior probability through Bayes' theorem, and outputs the initial state vector S0 = [x, v, T, f, h].

[0029] The high-precision S0 in step 1 reduces the initial error accumulation of the Transformer. RL uses this input to dynamically adjust the weights w1, w2, and w3, forming an enhancement effect of "high-dimensional calibration-deep prediction", which significantly extends the prediction window and improves robustness.

[0030] Furthermore, in this invention, step 2 is specifically described as follows: the dual-wavelength laser vision sensor, acoustic sensor, and infrared thermal imaging array of the multi-source sensing module provide the previous state S. t-1 The edge computing accelerator of the intelligent prediction module runs a Transformer model, processes high-dimensional inputs, and outputs predicted values. The embedded industrial computer runs an RL algorithm, calculates rewards, and optimizes parameters. The reinforcement learning reward function is: Ri t =w1·Q t +w2·(1-E t )-w3·D t ;

[0031] R t Total Reward; Q t Weld quality score, calculated from heatmap and point cloud computing; E t Energy consumption percentage, calculated by integrating current and voltage; D t : Defect probability, analyzed by acoustic anomaly; w1, w2, w3: Weights;

[0032] Transformer prediction output:

[0033] Y t Predicted output, weld width W t Melting depth M t Energy consumption E t Defect probability D t S t-1 : The state vector of the previous moment; A t Attention weight matrix; α i and V i These are the attention coefficient and the feature mapping value, respectively.

[0034] The multi-source sensing module provides S t-1 The FPGA runs a Transformer, which calculates Yt through an attention mechanism. The industrial computer then calculates Yt based on this information. t Calculate Q t E t D t Substitute into R t Optimize welding speed v through RL t Current I t Angle θ t Output the optimized parameter vector P t =[v t I t θ t ].

[0035] Step 2's Transformer attention mechanism focuses on the key features of S0, and RL optimizes the multi-objective function R through dynamic trial and error. t .

[0036] Predicting Y in step 2 t and parameter P t Feedforward control is provided for step 3, Kalman filtering and fuzzy control form feedback regulation, and force feedback and cooling system enhance execution accuracy. This multi-level synergy of "feedforward prediction-feedback regulation-physical execution" enables ultra-high frequency closed-loop control and adaptability to complex workpieces.

[0037] Furthermore, in this invention, step 3 specifically involves the following process: the dual-wavelength laser vision sensor, acoustic sensor, and infrared thermal imaging array of the multi-source sensing module provide real-time observation values ​​Z. t The embedded industrial computer in the intelligent prediction module runs Kalman filtering and fuzzy control algorithms; the six-axis welding robot in the dynamic execution module integrates force feedback sensors to adjust the path; the servo motor of the two-degree-of-freedom welding torch adjustment mechanism adjusts the height and tilt angle; and the adaptive power control unit adjusts the pulse frequency F. tThe microchannel cooling system uses solenoid valves to control the cooling flow rate C. t Kalman filter state update:

[0038] State estimation; Predicted state, from Y t ;K t Kalman gain, Z t H: Real-time observations;

[0039] Fuzzy control output: U t : Control the output pulse frequency F t Cooling flow rate C t Angle θ t ;e t Error, Y t -Z t ;Δe t Error rate of change; μ j u j These are the membership degree and rule outputs, respectively.

[0040] Six-axis robot along P t Path movement, multi-source sensing module collects Z t The industrial computer runs a Kalman filter and fuses the Y-axis. t and Z t renew According to error e t and Δe t Fuzzy control calculation U t The welding torch adjustment mechanism adjusts the angle, and the power control unit switches to F. t Cooling system adjustment C t Output real-time adjustment parameter U t =[F t C t θ t Kalman filtering smooths noise based on Bayesian estimation, fuzzy control optimizes execution through nonlinear mapping, and the hardware module implements high-frequency regulation based on physical principles.

[0041] Step 3: Kalman filtering based on state-space model fusion of Y t and Z t Fuzzy control utilizes membership functions based on the error e t =Y t -Z t Dynamic adjustment. Force feedback optimization path for a six-axis robot, dual-degree-of-freedom welding torch and pulsed power supply for stable welding based on molten pool dynamics, and microchannel cooling using Bernoulli's principle to control heat flow.

[0042] Step 3 updates the Transformer weights based on the least squares method and regularization using the loss function L, and Q-learning optimizes the strategy using reinforcement learning principles.

[0043] Furthermore, in this invention, step 4 is specifically described as follows: the dual-wavelength laser vision sensor, acoustic sensor, and infrared thermal imaging array of the multi-source sensing module provide the actual value Z. i The edge computing accelerator of the intelligent prediction module calculates the loss function and updates the Transformer weights. The embedded industrial computer runs Q-learning to update the RL policy. The loss function is:

[0044] L: Total loss; Y i : Predicted value; Z i : Actual value; N: Number of samples; λ: Regularization coefficient; W: Model weights;

[0045] RL Policy Update: Q(s) t ,a t )←Q(s t ,a t )+α·(R t +γ·maxQ(s t+1 ,a)-Q(s t ,a t ));

[0046] Q(s t a t ): State-action value function; α: Learning rate; R t : Instant reward; γ: Discount factor;

[0047] Multi-source sensing module acquires Z i The FPGA calculates L and updates W using gradient descent, while the industrial computer calculates W based on R. t And state transition, iteratively update Q(s) t a t ), output the updated model parameters and strategy.

[0048] Real-time data Z in step 3 t The execution results provide feedback for step 4, and the online updates of L and Q, in turn, optimize the prediction in step 2 and the control in step 3, forming a closed-loop adaptive system of "execution-feedback-evolution". This synergy enables the system to have self-learning ability and instantaneous responsiveness.

[0049] Furthermore, in this invention, steps 1 and 2 are connected by the initial state S0 and the predicted value Y. t The collaboration between the two methods enables a 20-minute prediction window; steps 2 and 3 utilize the prediction parameter P. tWith real-time adjustment U t Closed-loop coordination achieves a 0.5-second adjustment cycle; steps 3 and 4 utilize real-time data Z... t It achieves self-evolutionary stability through feedback with model updates.

[0050] Beneficial effects: The technical solution of this application has the following technical effects:

[0051] This invention significantly improves the overall performance of the welding process through the collaborative design of multi-source sensing, intelligent prediction, and dynamic execution modules, along with optimized fabrication methods. The system can accurately sense the welding environment and adaptively adjust its initial state, effectively addressing the challenges of complex working conditions and highly reflective materials. The intelligent prediction module achieves long-term trend prediction and multi-objective optimization through advanced algorithms, balancing the needs of quality, efficiency, and energy consumption. The dynamic execution module uses high-frequency closed-loop control to rapidly respond to instantaneous changes during the welding process, ensuring high consistency in weld quality and extremely low defect occurrence, while precise cooling control reduces the damage to workpiece performance caused by heat. Furthermore, the system possesses self-evolutionary capabilities, continuously optimizing models and strategies based on real-time feedback to maintain long-term operational stability. This comprehensive intelligent design not only improves welding efficiency and product quality but also enhances adaptability to complex workpieces and diverse production scenarios, reducing manual intervention and resource waste. It provides a highly efficient, reliable, and green welding solution for modern manufacturing, demonstrating significant technological advancements and practical value.

[0052] It should be understood that all combinations of the foregoing concepts and the additional concepts described in more detail below can be considered part of the inventive subject matter of this disclosure, provided that such concepts do not contradict each other.

[0053] The foregoing and other aspects, embodiments, and features of the teachings of the present invention will be more fully understood from the following description in conjunction with the accompanying drawings. Other additional aspects of the invention, such as features and / or beneficial effects of exemplary embodiments, will become apparent from the following description or may be learned through practice of specific embodiments according to the teachings of the present invention. Attached Figure Description

[0054] The accompanying drawings are not intended to be drawn to scale. In the drawings, each identical or nearly identical component shown in the various figures may be denoted by the same reference numeral. For clarity, not every component is labeled in each figure. Embodiments of various aspects of the invention will now be described by way of example and with reference to the accompanying drawings, wherein:

[0055] Figure 1 This is a schematic diagram of the system architecture of the present invention. Detailed Implementation

[0056] To better understand the technical content of this invention, specific embodiments are described below in conjunction with the accompanying drawings. Various aspects of the invention are described in this disclosure with reference to the accompanying drawings, which illustrate numerous illustrative embodiments. The embodiments of this disclosure are not necessarily defined to include all aspects of the invention. It should be understood that the various concepts and embodiments described above, as well as those described in more detail below, can be implemented in any of many ways, because the concepts and embodiments disclosed in this invention are not limited to any particular implementation. Furthermore, some aspects of this invention can be used alone or in any suitable combination with other aspects of this invention.

[0057] Example 1: This example provides a highly efficient intelligent welding system and its working method for butt welding of steel plates. The workpiece is made of Q235 material, with dimensions of 500mm × 300mm × 10mm and a weld length of 400mm. The system includes a multi-source sensing module, an intelligent prediction module, and a dynamic execution module, aiming to achieve a high-precision, high-efficiency, and high-quality welding process.

[0058] The multi-source sensing module is the core of the system's sensing capabilities. Its hardware includes: a dual-wavelength laser vision sensor (650nm red and 850nm near-infrared with an accuracy of 0.01mm, mounted above the welding torch); a multi-band acoustic sensor (microphone array with a frequency of 20Hz-20kHz and a sensitivity of -40dB, positioned on both sides of the welding area); an infrared thermal imaging array (0.1℃ resolution and 30° field of view, mounted above the workpiece); and an environmental adaptation unit (humidity sensor with an accuracy of ±2% and air pressure sensor with an accuracy of ±0.1kPa, integrated into the bracket).

[0059] A dual-wavelength laser vision sensor uses triangulation to overcome surface reflection interference from the steel plate by leveraging the complementarity of red and near-infrared light, generating a 3D point cloud of the weld location. A multi-band acoustic sensor acquires arc acoustic signals and extracts frequency features based on Fourier transform to reflect the dynamics of the molten pool. An infrared thermal imaging array measures temperature distribution according to the law of thermal radiation, generating a thermal map. An environmental adaptation unit monitors humidity and air pressure, correcting data deviations through a calibration model. An industrial computer fuses the data and runs a Bayesian model. Output the initial state S0 = [x, v, T, f, h].

[0060] The multi-source sensing module controls the initial weld position error to ±0.02mm, remaining stable even in extreme environments with 90% humidity or 85kPa air pressure, providing high-precision input data for subsequent prediction and execution, and improving the system's environmental adaptability.

[0061] The intelligent prediction module is responsible for trend prediction and parameter optimization. Its hardware includes: an edge computing accelerator: FPGA, Xilinx Zynq-7000, with a latency of <10ms, running a Transformer-LSTM model; and an embedded industrial computer: ARM Cortex-A9, with 8GB of memory, running a reinforcement learning algorithm.

[0062] FPGA receives S from multi-source sensing module t-1 Through the Transformer model Predict weld width, penetration depth, energy consumption, and defect probability over 20 minutes. An attention mechanism focuses on key features, while LSTM captures the temporal trend. An industrial computer runs a reinforcement learning reward function R. t =w1·Q t +w2·(1-E t )-w3·D t Weights w1 = 0.5, w2 = 0.3, w3 = 0.2, optimization parameter P t =[v t I t θ t ].

[0063] The intelligent prediction module achieves an ultra-long prediction window of 20 minutes, improving accuracy from 97% initially to 99% after 10 attempts. Optimized parameters increase welding speed to 1.5m / min and reduce energy consumption to 0.08kWh / m, achieving a balance between quality, efficiency, and energy consumption.

[0064] The dynamic execution module is responsible for welding execution and includes the following sub-modules:

[0065] Six-axis welding robot: The six-axis robotic arm includes 6 servo motors (200W, range 1.5m, load 10kg), force feedback sensors (accuracy ±0.1N, end effector mounted), motion controller (ARM Cortex-M7, running Q-learning), and optical encoders (resolution 0.01°, 1 per axis).

[0066] Dual-degree-of-freedom welding torch adjustment mechanism: dual-axis servo motor (height 0-50mm, tilt ±15°, response 0.1 seconds), linear guide rail (accuracy 0.01mm), angle sensor (accuracy ±0.1°), servo driver (stepper controller).

[0067] Adaptive power control unit: IGBT inverter (50-300A, 15-30V), pulse generator (TI C2000DSP, 10-500Hz), current sensor (accuracy ±0.5A), control circuit (STM32).

[0068] Microchannel cooling system: microchannel nozzle (0.1mm orifice, injecting argon gas), solenoid valve (0-10L / min, response time 0.05 seconds), gas flow meter (accuracy ±0.1L / min), pressure regulator (0.2-0.5MPa).

[0069] The six-axis welding robot uses a motion controller based on P t The path is planned, the force feedback sensor detects the resistance, the encoder provides angle feedback, and the servo motor adjusts the position. The two-degree-of-freedom welding torch adjustment mechanism receives U... t θ t The motor adjusts its height and angle, and the sensor provides feedback. The control circuit of the adaptive power control unit adjusts according to U... t The Ft adjustment pulse in the inverter power supply outputs It, and a current sensor provides closed-loop control. The microchannel cooling system adjusts the injection based on Ct via a solenoid valve, with flow meter feedback and a pressure regulator stabilizing the airflow. An industrial computer runs a Kalman filter. And fuzzy control U t .

[0070] The dynamic execution module achieves a 0.5-second adjustment cycle, weld consistency reaches 99.8%, defect rate is as low as 0.2%, and heat-affected zone is controlled within 1.5mm, significantly improving welding quality and efficiency.

[0071] The multi-source sensing module and the intelligent prediction module work together to achieve an ultra-long prediction window (20 minutes) and high accuracy (99%), compared to traditional methods which only take 5 minutes and achieve 85%. Multi-source sensing provides high-dimensional data (point cloud, sound spectrum, heat map), the Bayesian model reduces initial uncertainty, the Transformer captures the spatiotemporal correlation of multi-dimensional features through the attention mechanism, and reinforcement learning optimizes the trade-offs between multiple objectives, breaking through the limitations of single-data prediction.

[0072] The intelligent prediction module and the dynamic execution module work together to achieve high-frequency adjustment of 0.5 seconds and 99.8% consistency, compared to 10-30 seconds and 96.7% consistency for traditional methods. Prediction parameter P t It provides feedforward guidance, Kalman filtering to fuse real-time data, fuzzy control for dynamic adjustment, force feedback and pulse modulation to optimize execution, and closed-loop collaboration between prediction and execution to achieve ultra-fast response and quality improvement.

[0073] The dynamic execution module, in collaboration with the multi-source sensing module and the intelligent prediction module, reduces volatility to 0.1%, compared to 2%-5% for traditional methods. The real-time data Zt from the dynamic execution module is updated through multi-source sensing feedback, loss function (L), and Q-learning. High-frequency feedback supported by hardware enhances evolutionary efficiency, and the execution-sensing-prediction loop forms adaptive learning capability.

[0074] The overall working principle of the system is a highly collaborative closed-loop optimization process. First, the multi-source sensing module starts, using a dual-wavelength laser vision sensor to scan the workpiece and generate a point cloud, a multi-band acoustic sensor to record the initial arc sound, an infrared thermal imaging array to create a heat map, an environmental adaptation unit to calibrate humidity and air pressure, and an industrial computer to run a Bayesian model to generate the initial state S0. Next, the intelligent prediction module takes over, with the FPGA running a Transformer model to predict the 20-minute trend, and the industrial computer optimizing parameters Pt through reinforcement learning. Subsequently, the dynamic execution module executes welding according to Pt: the six-axis welding robot adjusts its path, the two-degree-of-freedom welding torch adjusts its angle, the adaptive power supply controls the pulses, the microchannel cooling system sprays gas, the multi-source sensing module collects Zt in real time, and the industrial computer generates an adjustment signal U through Kalman filtering and fuzzy control. t Finally, the actual data Z of the dynamic execution module. i Feedback is sent to the intelligent prediction module, where the FPGA's loss function updates the Transformer. The industrial computer then uses a Q-learning optimization strategy to form a self-evolving loop. The entire process achieves full-process intelligence from initialization to long-term operation through high-precision input from multi-source sensing, forward-looking guidance from intelligent prediction, rapid response from dynamic execution, and adaptive optimization from feedback.

[0075] The results showed that the weld width was 5.0 mm (±0.02 mm), the penetration depth was 8 mm, the defect rate was 0.2%, the energy consumption was 0.08 kWh / m, and the fluctuation rate dropped to 0.1% after 10 welding cycles, which verified the superior performance of the system.

[0076] The following experiments were designed to verify the effectiveness of the optimized preparation method in terms of prediction accuracy, welding quality consistency, defect rate, adjustment cycle, and long-term stability. The experiments were also compared with traditional methods to demonstrate its superior technical performance.

[0077] Experimental subjects

[0078] The method of this invention, based on optimized steps, includes multi-dimensional initialization, AI multi-objective prediction, adaptive closed-loop execution, real-time feedback and self-evolution, combined with hardware modules such as a multi-source perception module, an intelligent prediction module, and a dynamic execution module.

[0079] The comparison is a traditional method, using a semi-automatic welding system with conventional PID control, which only uses a single optical sensor and fixed parameter adjustment, without AI prediction and online learning functions.

[0080] Experimental conditions

[0081] The workpiece is an aluminum alloy plate, 500mm×300mm×5mm, material 6061-T6.

[0082] Welding type: butt weld, weld length 400mm.

[0083] The environmental conditions are as follows: normal environment is 25℃ temperature, 50% humidity, and 101kPa air pressure. Extreme environment is -10℃ temperature, 90% humidity, and 85kPa air pressure (simulating high altitude).

[0084] The equipment includes:

[0085] This invention includes: a dual-wavelength laser vision sensor (accuracy 0.01mm), a multi-band acoustic wave sensor (20Hz-20kHz), an infrared thermal imaging array (resolution 0.1℃), a six-axis robot (force feedback accuracy ±0.1N), an FPGA accelerator, an industrial computer, etc.

[0086] Comparative examples: traditional optical sensor (accuracy 0.1mm), three-axis welding robot, fixed power supply (no pulse modulation).

[0087] Experimental steps

[0088] During the preparation stage, clean the surface of the workpiece to ensure that there is no oil or oxide layer.

[0089] This invention involves installing a multi-source sensing module (laser, sound wave, thermal imaging, environmental sensor) in a welding workstation, and connecting it to an intelligent prediction module (FPGA and industrial computer) and a dynamic execution module (robot, welding torch, power supply, cooling system).

[0090] Comparative example: A single optical sensor and a three-axis robot are installed, with fixed parameters set: speed 1m / min, current 180A, and voltage 22V.

[0091] Step 1: Multidimensional Initialization and Environmental Adaptive Calibration

[0092] This invention activates a multi-source sensing module: a laser sensor scans and generates a point cloud, an acoustic sensor records the initial arc sound, a thermal imaging array creates a heat map, and environmental sensors measure humidity and air pressure. An industrial computer runs a Bayesian model (P(S|E)) and outputs the initial state S0. Initial errors and weld position deviations are recorded.

[0093] A comparative optical sensor scanned the weld seam, with fixed parameters manually input and no environmental calibration. Initial errors were recorded.

[0094] Step 2: AI-driven multi-objective prediction and optimization planning

[0095] The present invention uses an FPGA to run a Transformer model, taking input S0, to predict the weld width W within 20 minutes. t Melting depth M t Energy consumption E t Defect probability D tIndustrial computer computing RL reward R t Optimize parameter P t =[v t I t θ t Record the prediction window and accuracy, and compare them with the actual values.

[0096] Comparative example: No prediction function, executes directly using fixed parameters.

[0097] Step 3: Adaptive Closed-Loop Welding Execution

[0098] The six-axis robot of the present invention along P t Path welding, multi-source sensing module collects real-time data Z t Industrial computers run Kalman filtering. And fuzzy control U t Adjust the pulse frequency F t Cooling flow rate C t Angle θ t The cycle time is 0.5 seconds. Record the adjustment cycle, consistency, and defect rate.

[0099] A comparative three-axis robot welds along a fixed path, controlled by PID, with a cycle of 15 seconds, and the same performance indicators are recorded.

[0100] Step 4: Real-time feedback and AI model self-evolution

[0101] The multi-source sensing module of this invention collects the actual value Z. i FPGA calculates the loss L to update the Transformer, while industrial computers update the RL policy Q(s). t a t Repeat the welding process 10 times and record the changes in prediction accuracy and quality fluctuation rate.

[0102] The comparison showed no feedback and self-evolution; the welding was repeated 10 times, and the same indicators were recorded.

[0103] Data acquisition and analysis are performed using high-precision measuring instruments, such as laser profilometers or microscopes, to detect weld width, penetration depth, and the number of defects. Consistency, defect rate, and energy consumption are calculated.

[0104] A table was created by comparing the results of the present invention with those of the comparative example.

[0105] Table 1: Performance Comparison Under Normal Environments

[0106] index This invention Comparative Example Initial error (mm) ±0.02 ±0.10 Forecast window (minutes) 20 none Prediction accuracy (%) 97 none Adjustment period (seconds) 0.5 15 Welding speed (m / min) 1.5 1.0 Weld width consistency (mm) 4.8±0.02(99.8％) 4.5±0.15(96.7％) Penetration depth (mm) 3.5±0.03 3.2±0.20 Defect rate (%) 0.2 1.5 Energy consumption (kWh / m³) 0.08 0.10

[0107] Table 2: Performance Comparison Under Extreme Environments

[0108] index This invention Comparative Example Initial error (mm) ±0.03 ±0.25 Prediction accuracy (%) 95 none Adjustment period (seconds) 0.5 20 Weld width consistency (mm) 4.8±0.03(99.4％) 4.5±0.30(93.3％) Defect rate (%) 0.3 2.5

[0109] Table 3: Self-evolution effect after 10 welding cycles

[0110] index This invention (first time) This invention (after 10 uses) Comparative Example (Initial) Comparative example (after 10 trials) Prediction accuracy (%) 95 99 none none Quality volatility (%) 0.5 0.1 2.0 2.2 Defect rate (%) 0.2 0.1 1.5 1.6

[0111] The comparative conclusions are as follows: This invention achieves an ultra-long prediction window of 20 minutes, with an initial accuracy of 97%, which improves to 99% after 10 trials; the comparative example has no prediction function.

[0112] This invention significantly improves prediction performance through the synergy of steps 1 and 2. The invention has an adjustment cycle of 0.5 seconds, achieving 99.8% consistency and a defect rate of 0.2%; the comparative example has a cycle of 15-20 seconds, achieving 93.3%-96.7% consistency and a defect rate of 1.5%-2.5%. The high-frequency closed-loop synergy of steps 2 and 3 achieves ultra-high accuracy and a low defect rate.

[0113] After 10 welding cycles, the prediction accuracy of this invention is improved to 99%, and the quality fluctuation rate is reduced to 0.1%; the comparative example has no self-evolution and the fluctuation rate is 2.0%-2.2%. The feedback evolution synergy between steps 3 and 4 ensures long-term stability.

[0114] This invention maintains 95% prediction accuracy and a 0.3% defect rate even under extreme conditions; the comparative model shows a significant decrease in performance, with a defect rate of 2.5%. The environmental calibration in step 1, in conjunction with subsequent steps, enhances robustness.

[0115] In principle, the Bayesian model in steps 1 and 2 reduces initial uncertainty through data fusion from multi-source sensing modules, while the Transformer and RL utilize high-precision S0 to capture spatiotemporal features and multi-objective tradeoffs. The ultra-long prediction window stems from the extremely low starting point of the initial error and the deep modeling of AI.

[0116] Predicted values ​​Y in steps 2 and 3 t and parameter P t The dynamic execution module provides feedforward for Kalman filtering and fuzzy control, and its hardware implements high-frequency response. A 0.5-second settling period and 99.8% consistency stem from the closed-loop coordination of prediction and execution. Real-time data Z from steps 3 and 4. t The model is updated using the L- and Q-learning loss functions, and hardware-supported high-frequency feedback enhances evolutionary efficiency. Self-evolution improves accuracy to 99% with a volatility of 0.1%, stemming from the dynamic coupling of execution and learning. Experimental data show that this invention significantly outperforms comparative methods in terms of prediction window, adjustment speed, quality consistency, and long-term stability, demonstrating a breakthrough in step-by-step synergy.

[0117] While the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Those skilled in the art can make various modifications and refinements without departing from the spirit and scope of the invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. A highly efficient intelligent welding system, characterized in that, Includes the following modules: The multi-source sensing module includes a dual-wavelength laser vision sensor, a multi-band acoustic sensor, an infrared thermal imaging array, and an environmental adaptation unit, which are used to acquire weld position, arc sound signal, temperature distribution, and environmental parameters in real time. The intelligent prediction module, including an edge computing accelerator and an embedded industrial computer, incorporates a prediction model that fuses Transformer and Long Short-Term Memory networks with a reinforcement learning optimization algorithm to predict weld quality, energy consumption, and defect probability, and generate optimized parameters. The dynamic execution module includes a six-axis welding robot, a two-degree-of-freedom welding torch adjustment mechanism, an adaptive power control unit, and a microchannel cooling system, which are used to dynamically adjust the welding path, power parameters, and cooling intensity based on prediction results. Among them, the multi-source sensing module and the intelligent prediction module achieve high-precision initialization through data fusion, the intelligent prediction module and the dynamic execution module achieve high-frequency adjustment through closed-loop control, and the dynamic execution module and the intelligent prediction module achieve model self-evolution through real-time feedback; The execution steps are as follows: Step 1: Multidimensional initialization and environmental adaptive calibration. Initial data is collected using the multi-source sensing module, and the Bayesian probability model is run through the industrial computer in the intelligent prediction module to calibrate environmental disturbances and generate a high-precision initial state vector. Step 2: AI-driven multi-objective prediction and optimization planning. The edge computing accelerator in the intelligent prediction module is used to run the Transformer model to predict welding trends. The industrial computer performs reinforcement learning to optimize parameters and generate a multi-objective planning scheme. Step 3: Adaptive closed-loop welding execution. The dynamic execution module performs welding according to the predicted parameters, the multi-source sensing module monitors in real time, and the intelligent prediction module runs Kalman filtering and fuzzy control to dynamically adjust the execution parameters. Step 4: Real-time feedback and AI model self-evolution. The multi-source perception module provides actual data, and the intelligent prediction module updates the Transformer model and RL strategy through edge computing accelerators and industrial computers to improve long-term performance. The specific process of step 1 is as follows: The dual-wavelength laser vision sensor of the multi-source sensing module scans the workpiece surface to generate a three-dimensional point cloud of the weld position x; the multi-band acoustic sensor records the initial arc sound frequency f through a microphone array; the infrared thermal imaging array measures the workpiece temperature distribution T and generates a two-dimensional thermal map; the environmental adaptation unit includes a humidity sensor and a pressure sensor to collect environmental parameters E, including humidity h and pressure v; the prediction module includes an embedded industrial computer that runs a Bayesian algorithm, stores historical data, and calculates the probability. The calculation formula is: ; P(S|E): Posterior probability, representing the welding state (S) under a given environmental parameter (E); P(E|S): Likelihood probability, generated from sensor data; P(S): Prior probability, based on historical welding data; P(E): Marginal probability of environmental parameter. A dual-wavelength laser vision sensor scans the workpiece and outputs point cloud data; an acoustic sensor records the arc sound spectrum; an infrared thermal imaging array generates a thermal map; an environmental adaptation unit measures humidity and air pressure; an industrial computer receives this data, calculates a likelihood model based on sensor accuracy and historical database, outputs the posterior probability through Bayes' theorem, and outputs the initial state vector S0=[x, v, T, f, h]; The specific process of step 2 shown is as follows: the dual-wavelength laser vision sensor, acoustic sensor, and infrared thermal imaging array of the multi-source sensing module provide the previous state S. t-1 The edge computing accelerator of the intelligent prediction module runs a Transformer model, processes high-dimensional inputs, and outputs predicted values. The embedded industrial computer runs an RL algorithm, calculates rewards, and optimizes parameters. The reinforcement learning reward function is as follows: ; R t Total Reward; Q t Weld quality score, calculated from heatmap and point cloud computing; E t Energy consumption percentage, calculated by integrating current and voltage; D t : Defect probability, analyzed by acoustic anomaly; w1, w2, w3: Weights; Transformer prediction output: ; Y t Predicted outputs: weld width, penetration depth, and energy consumption E. t Defect probability D t S t-1 : The state vector of the previous moment; A t Attention weight matrix; α i and V i These are the attention coefficient and the feature mapping value, respectively. The multi-source sensing module provides S t-1 The FPGA runs a Transformer, which calculates Yt through an attention mechanism. The industrial computer then calculates Yt based on this information. t Calculate Q t E t D t Substitute into R t Optimize welding speed v through RL t Current I t Angle θ t Output the optimized parameter vector P t =[v t I t θ t ].

2. The efficient intelligent welding system according to claim 1, characterized in that, The dynamic execution module includes: The six-axis welding robot includes a six-axis robotic arm driven by six servo motors, a force feedback sensor mounted at the end effector, an embedded motion controller, and a high-precision optical encoder. The force feedback sensor has an accuracy of ±0.1N, and the encoder has a resolution of 0.01°, which are used to dynamically adjust the welding path based on predicted parameters. The dual-degree-of-freedom welding torch adjustment mechanism includes a dual-axis servo motor, a linear guide rail, an angle sensor, and a servo driver. The dual-axis servo motor adjusts the height and tilt angle with an adjustment cycle of 0.5 seconds to ensure weld consistency. The adaptive power control unit includes an IGBT inverter power supply, a pulse generator, a current sensor, and a control circuit. The inverter power supply has an output range of 50-300A, the pulse generator supports 10-500Hz modulation, and the current sensor has an accuracy of ±0.5A. It is used to dynamically adjust the current and pulse frequency to optimize the stability of the molten pool. The microchannel cooling system includes a microchannel nozzle, a solenoid valve, a gas flow meter, and a pressure regulator. The microchannel nozzle has an orifice diameter of 0.1 mm, the solenoid valve adjusts the flow rate range to 0-10 L / min, the flow meter has an accuracy of ±0.1 L / min, and the pressure regulator maintains a pressure of 0.2-0.5 MPa to control the heat-affected zone.

3. The high-efficiency intelligent welding system according to claim 2, characterized in that, The force feedback sensor and motion controller of the six-axis welding robot work together to optimize the path by sensing the contact force of the workpiece in real time and running a reinforcement learning algorithm; the dual-axis servo motor and angle sensor of the dual-degree-of-freedom welding torch adjustment mechanism achieve closed-loop adjustment through a servo driver; the pulse generator and current sensor of the adaptive power control unit dynamically adjust the pulse waveform through a control circuit; the solenoid valve and gas flow meter of the microchannel cooling system precisely adjust the cooling flow through feedback control.

4. The efficient intelligent welding system according to claim 2, characterized in that, The six-axis welding robot of the dynamic execution module adjusts its path according to the optimized parameters output by the intelligent prediction module, and the dual-degree-of-freedom welding torch adjustment mechanism, adaptive power control unit and microchannel cooling system adjust the angle, pulse frequency and cooling flow rate according to the fuzzy control output.

5. The efficient intelligent welding system according to claim 1, characterized in that, The specific process of step 3 is as follows: the dual-wavelength laser vision sensor, acoustic sensor, and infrared thermal imaging array of the multi-source sensing module provide real-time observation values ​​Z. t The embedded industrial computer in the intelligent prediction module runs Kalman filtering and fuzzy control algorithms; the six-axis welding robot in the dynamic execution module integrates force feedback sensors to adjust the path; the servo motor of the two-degree-of-freedom welding torch adjustment mechanism adjusts the height and tilt angle; and the adaptive power control unit adjusts the pulse frequency F. t The microchannel cooling system uses solenoid valves to control the cooling flow rate C. t Kalman filter state update: ; State estimation; Predicted state, from Y t ;K t Kalman gain, Z t H: Real-time observations; Fuzzy control output: U t : Control the output pulse frequency F t Cooling flow rate C t Angle θ t ;e t Error, Y t -Z t ;Δe t Error rate of change; μ j u j These are the membership degree and rule outputs, respectively. Six-axis robot along P t Path movement, multi-source sensing module collects Z t The industrial computer runs a Kalman filter and fuses the Y-axis. t and Z t renew According to error e t and Δe t Fuzzy control calculation U t The welding torch adjustment mechanism adjusts the angle, and the power control unit switches to F. t Cooling system adjustment C t Output real-time adjustment parameter U t =[F t C t θ t ].

6. The efficient intelligent welding system according to claim 1, characterized in that, The specific process of step 4 is as follows: the dual-wavelength laser vision sensor, acoustic sensor, and infrared thermal imaging array of the multi-source sensing module provide the actual value Z. i The edge computing accelerator of the intelligent prediction module calculates the loss function and updates the Transformer weights. The embedded industrial computer runs Q-learning to update the RL policy. The loss function is: ; L: Total loss; Y i : Predicted value; Z i Actual value; N: Number of samples; λ: Regularization coefficient; W: Model weights; RL policy update: ; Q(s) t a t ): State-action value function; α: learning rate; R t Instant rewards; γ: Discount factor; Multi-source sensing module acquires Z i The FPGA calculates L and updates W using gradient descent, while the industrial computer calculates W based on R. t And state transition, iteratively update Q(s) t a t ), outputting the updated model parameters and strategy.

7. The efficient intelligent welding system according to claim 6, characterized in that, Steps 1 and 2 are based on the initial state S0 and the predicted value Y. t The collaboration between the two methods enables a 20-minute prediction window; steps 2 and 3 utilize the prediction parameter P. t With real-time adjustment U t Closed-loop coordination achieves a 0.5-second adjustment cycle; steps 3 and 4 utilize real-time data Z... t It achieves self-evolutionary stability through feedback with model updates.

Images